EHRs with coded clinical data better predict general vs. specific antidepressant treatment response
Coded clinical data in electronic health records appeared to help predict general antidepressant treatment response but not specific medication response, according to results of a retrospective cohort study published in JAMA Network Open.
“A key challenge in [prior] studies has been the paucity of head-to-head antidepressant studies distinguishing factors associated with poor outcomes overall from factors associated with poor outcomes specific to a given medication,” Michael C. Hughes, PhD, of the department of computer science at Tufts University, and colleagues wrote. “Traditional tests of interaction compound this problem because they are best powered for opposing associations (ie, markers associated with better outcome in one group and poorer outcome in another), when in reality, this may not comport with biologic characteristics. Furthermore, even in head-to-head studies, there are rarely replication cohorts to follow up initial associations.”
Researchers have used EHR or administrative data sets to assess clinical outcomes in other settings. Doing so provided sufficiently large real-world cohorts for identifying and validating predictors, Hughes and colleagues noted. Further, these data sets operate on data already available at the point of care, which allows for clinical adoption without the use of new rating scales or measures.
In the current study, the investigators sought to evaluate a model using EHRs to identify treatment response predictors among patients with major depressive disorder. They included data of 81,630 adults with a coded MDD diagnosis from two academic medical centers in Boston across 2 decades. Treatment with at least one of 11 standard antidepressants served as the exposure. The main outcome was stable treatment response, intended as a proxy for treatment effectiveness, defined as continued antidepressant prescription for 90 days.
Hughes and colleagues extracted 10 interpretable covariates from coded clinical data for stability prediction using supervised topic models. They used data from one hospital system, which was site A, to train generalized linear models and ensembles of decision trees to predict stability outcomes from topic features that summarize patient history, according to the study. They evaluated held-out patients from site A, as well as individuals from a second hospital system, which was site B.
Results showed 55,303 patients achieved a stable treatment regimen response during follow up. The mean area under the receiver operating characteristic curve (AUC) for discrimination of the general stability outcome for held-out patients from site A was 0.627 (95% CI, 0.615-0.639) for the supervised topic model with 10 covariates. The AUC for site B was 0.619 (95% CI, 0.61-0.627). Models that predicted stability specific to a particular drug were not associated with improvements in the prediction of general stability (specific AUC = 0.647; 95% CI, 0.635-0.658; general AUC = 0.661; 95% CI, 0.648-0.672). The investigators observed that topics coherently captured clinical concepts associated with treatment response.
“The findings further suggest that features derived from supervised topic models provide more interpretable insights compared with raw coded features,” Hughes and colleagues wrote. “Although greater discrimination is likely required for clinical application, the results provide a transparent baseline for such studies.” – by Joe Gramigna
Disclosures: Hughes reports grants from Oracle during the conduct of the study. Please see the study for all other authors’ relevant financial disclosures.